Análise comparativa da classificação de séries temporais representadas como grafos

Abstract

Time series classification is a fundamental problem in various domains, such as medicine, finance, and engineering. Recently, graph-based approaches have shown promise for this task, enabling the capture of complex patterns and relationships in temporal data. This work presents a comparative study of different models for time series representation as graphs, investigating: (1) the main models available; (2) how they compare in terms of agreement and performance; (3) their effectiveness compared to convolutional models; (4) statistically significant differences among models; (5) the most effective architectures; and (6) the influence of time series characteristics on classifier performance. Experiments were conducted on 16 datasets, including 10 from the UCR Archive and 6 derived from CODE-15\%, a repository of electrocardiograms. The characterization of these datasets revealed that trend strength, seasonality, noise, and autocorrelation are key factors in model performance. The results, evaluated through Friedman statistical tests with Nemenyi post-hoc analysis and agreement assessment via Fleiss' Kappa, showed that the evaluated models exhibit statistically similar performance and that there is a slight to moderate agreement between models ($\kappa$ = 0.0176), with greater stability for structured time series and lower agreement in noisy datasets. The choice of the best model depends on the type of time series: ROCKET performed best for ECG and spectral datasets, SimTSC-GNA was more effective for highly autocorrelated series, and VisibilityGraph-3feat outperformed others in seasonal time series, as detailed in the model recommendations. This study establishes the first comprehensive statistical analysis of different CSTG (Classification of Time Series as Graphs) models, providing relevant insights into their applicability and guiding future research on graph-based learning for time series classification.